Bayesian Optimization-Based Machine Learning Inversion of 237NpTransmutation Depletion in ADS

The Accelerator Driven Advanced Nuclear Energy System (ADANES) is a novel closed nuclear fuel cycle system consisting of an Accelerator Driven Sub-critical System (ADS) and a spent fuel recycling system. Its primary function is the transmutation of long-lived minor actinides. To monitor the depletion of 237 Np in the transmutation fuel rod of an ADS reactor core, this paper proposes an offline monitoring scheme based on a Bayesian-optimized machine learning model that integrate nuclide characteristic peak areas with operational parameters. The scheme establishes a nonlinear mapping relationship among 237 Np depletion, characteristic peak area counts of key nuclides, and known operational histories using a database generated from simulations of an ex-core measurement system equipped with a high-purity germanium (HPGe) detector, thereby enabling rapid inversion of the 237 Np transmutation state. Inversion results obtained from six Bayesian-optimized machine learning models demonstrate that the BO-XGBoost model achieves the best performance, yielding a mean relative error (MRE) of 1.47% and a maximum relative error (MaxRE) of 5.32%. Analysis of nuclide and operational parameter feature configurations reveals that burnup time is the dominant operational parameter governing prediction accuracy. When critical operational history information such as burnup history is unavailable, gamma-ray characteristic peak areas can effectively compensate for the deficiency in operational parameters and enhance prediction accuracy, underscoring the gamma-spectroscopic information as an essential source for high-precision inversion. Among the nuclide combinations examined, N3 ( 125 Sb, 134 Cs, 144 Ce, 154 Eu) exhibits the most favorable prediction stability.